Re: DoS attack against TCP services

[6bone%6bone.informatik.uni-leipzig.de@localhost]

On Wed, 4 Feb 2015, Sverre Froyen wrote:

> > I'd also look at the open descriptors of the named process (although they
> > should be closed at this time, since TIME_WAIT means closed on this side,
> > and waiting for the 4 minutes to expire before killing the connection)...
> >
> > Also I'd record that information every minute or so to see how many
> > connections are added and how many are going away.
> >
> > Perhaps there is some bug triggered in the tcp stack and somehow connections
> > are not being GC’ed?
>
> This is vaguely similar to a problem I have seen from time to time. On my servers, it is usually port 80 that gets attacked. Someone opens TCP connections to this port on the server, sends no request, and leaves the connection open indefinitely. See http://mail-index.netbsd.org/netbsd-users/2011/01/04/msg007484.html
>
> When I test such a scenario to port 53 (using telnet), the connection shows as ESTABLISHED for 30 seconds. Then, presumably, named times-out and closes the connection. At this point netstat shows the connection as TIME_WAIT for another 10 seconds. After that it disappears.
>
> If I disable the network connection during the 30 second period before named times out, however, I instead observe the connection in FIN_WAIT_1 mode for another 30 minutes or so.
>
> This is on netbsd-6. I notice that your netstat output has the client and server columns in the reverse order from what I see. Could it be that in your netstat output, FIN_WAIT_1 is reported as TIME_WAIT?
>
> Regards,
> Sverre

The two problems are not identical. In my case, the connections are really 
in the TIME_WAIT state. Christos has also found that the 2MSL timer each 
connection is negative. If this value is negative, the connection should 
be removed.

The callout code in kern_timeout.c:

                if (delta <0)
                         cc->cc_ev_late.ev_count++;

At the same time, the problem occurs that expired entries are not deleted 
from the ndp table. 'ndp -a' shows expired entries.

Both problems occur only after several days of uptime. They probably have 
the same cause.

The problem you described is different.


Regards
Uwe